training_config_source.yaml

# @package _global_
# REFERENCE COPY — canonical is in this workstream's local lerobot clone at:
# ./lerobot/lerobot/configs/policy/act_diffusion_aloha_solo_real.yaml

# Single-arm (LEFT) ALOHA — Hybrid ACT+Diffusion policy.
# ACT encoder (ResNet18 + transformer) → DDIM diffusion U-Net → action chunks.
# 2 cameras: cam_left_wrist + cam_high.  state_dim=action_dim=9.
# DOE winner: batch=24, lr=3e-5 (2026-04-20)

seed: 1000
dataset_repo_id: JHeisler/aloha_solo_left_4_6_26

override_dataset_stats:
  observation.images.cam_left_wrist:
    mean: [[[0.485]], [[0.456]], [[0.406]]]
    std: [[[0.229]], [[0.224]], [[0.225]]]
  observation.images.cam_high:
    mean: [[[0.485]], [[0.456]], [[0.406]]]
    std: [[[0.229]], [[0.224]], [[0.225]]]

use_amp: true
use_torch_compile: true

training:
  offline_steps: 40000
  online_steps: 0
  eval_freq: -1
  save_freq: 10000
  log_freq: 100
  save_checkpoint: true

  batch_size: 28
  lr: 3.5e-5
  lr_backbone: 3.5e-5
  lr_warmup_steps: 500
  drop_n_last_frames: 2
  weight_decay: 1e-4
  grad_clip_norm: 10
  online_steps_between_rollouts: 1

  delta_timestamps:
    action: "[i / ${fps} for i in range(${policy.chunk_size})]"

eval:
  n_episodes: 50
  batch_size: 50

policy:
  name: hybrid_act_diffusion

  n_obs_steps: 1
  chunk_size: 100
  n_action_steps: 100

  input_shapes:
    observation.images.cam_left_wrist: [3, 480, 640]
    observation.images.cam_high: [3, 480, 640]
    observation.state: ["${env.state_dim}"]
  output_shapes:
    action: ["${env.action_dim}"]

  input_normalization_modes:
    observation.images.cam_left_wrist: mean_std
    observation.images.cam_high: mean_std
    observation.state: mean_std
  output_normalization_modes:
    action: mean_std

  # ACT visual encoder
  vision_backbone: resnet18
  pretrained_backbone_weights: ResNet18_Weights.IMAGENET1K_V1
  replace_final_stride_with_dilation: false
  pre_norm: false
  dim_model: 512
  n_heads: 8
  dim_feedforward: 3200
  feedforward_activation: relu
  n_encoder_layers: 4
  dropout: 0.1

  # Diffusion U-Net
  down_dims: [256, 512]
  kernel_size: 5
  n_groups: 8
  diffusion_step_embed_dim: 128
  use_film_scale_modulation: true

  # Noise scheduler
  noise_scheduler_type: DDPM
  num_train_timesteps: 100
  beta_schedule: squaredcos_cap_v2
  beta_start: 0.0001
  beta_end: 0.02
  prediction_type: epsilon
  clip_sample: true
  clip_sample_range: 1.0
  num_inference_steps: 10
  do_mask_loss_for_padding: true